Minimum detectable effect size for main outcomes (accounting for sample
design and clustering)
In our previous study involving standardized patients, community pharmacies provided antibiotics without a physician prescription for pediatric diarrhea and URTI in 60 percent of the cases (average of diarrhea and URTI). In absence of prior data on antibiotics consumption at the household level, we assume that this proportion reflects household-level consumption of non-prescribed antibiotics. A 20 percentage point reduction in this proportion would be meaningful for policy, so we would like to be able to detect this difference at the 5% significance level with at least 80% power. There is no reason to believe that the intervention would have a negative effect—i.e., raise antibiotic consumption—so we base our calculation on a one-tail test. Our intervention will vary at the ward level to avoid contamination, with 20 households selected per ward. We assume a conservative intra-cluster correlation of 0.2. With these assumptions, we need a total of 760 households (380 households in each arm) from 40 wards (20 in each arm) for the final analysis. Our prior experience suggests that non-response and attrition are very low in this setting (in the vicinity of 10%). Nonetheless, we plan on a 20% attrition rate as the study requires participants to use a phone App. Therefore, we will recruit and enroll 480 households in each arm, for a total of 960 households in the study.